NVIDIA DGX GH200 price: what to verify before budgeting

NVIDIA does not list a public DGX GH200 purchase price on the product page reviewed on October 16, 2026. The page directs buyers to contact NVIDIA. A useful budget therefore starts with a configuration-specific quote, not an unsourced price range copied from a different DGX server.

The name is a common source of confusion. A GH200 Grace Hopper Superchip, a server containing GH200 hardware, and the DGX GH200 supercomputer are different purchases. A cloud rate for one accelerator also does not price an entire interconnected system. This guide explains what to verify in a quote and how to compare ownership with rental for the workload you actually need to run.

What does the DGX GH200 price refer to?

NVIDIA announced DGX GH200 in May 2023 as a system connecting 256 Grace Hopper Superchips, with 144 TB of shared memory. Treat those as specifications of the announced architecture, not confirmation that an identical system is available to order today. Ask the supplier to identify the exact configuration, support status, and delivery terms.

Product name	What it describes	Pricing question
GH200 Grace Hopper Superchip	A Grace CPU and Hopper GPU joined through NVLink-C2C.	Does the quote cover a component, a complete server, or a cloud allocation?
DGX GH200	A large system connecting Grace Hopper Superchips through an NVLink Switch System.	Which nodes, switches, racks, software, and services are included?
DGX H100 or DGX H200	Integrated eight-GPU servers using H100 or H200 accelerators.	Which DGX model and bill of materials does the supplier actually name?
GH200 cloud instance	A provider-defined allocation using GH200 hardware.	How many Superchips, how much memory, and which network and usage terms does the rate cover?

NVIDIA's GH200 product information and DGX H100/H200 documentation describe the different architectures. A seller's price for an eight-GPU DGX H100 should not be relabeled as a DGX GH200 price. Our DGX A100 guide covers the earlier eight-GPU Ampere system separately.

Why the memory and interconnect matter

The original DGX GH200 design combined 480 GB of CPU memory and 96 GB of HBM3 GPU memory per Superchip. Its 144 TB headline therefore does not mean 144 TB of GPU HBM. NVIDIA's architecture explanation describes how GPUs access memory across the connected system.

NVLink-C2C connects the CPU and GPU within a Superchip; the NVLink Switch System connects Superchips at the system level. Those are separate parts of the design. Later GH200 configurations also use HBM3e, so confirm the memory variant rather than applying one specification to every GH200 listing.

For procurement, the question is whether the application benefits from that memory access and communication topology. Test the required precision, dataset, model, and software environment. Total memory and peak compute figures alone do not predict usable throughput. Our HPC simulation guide explains why communication and serial work can limit scaling.

What a usable DGX GH200 quote should include

The DGX GH200 product page provides a contact route rather than a public price. Request an itemized proposal with a date, currency, validity period, and explicit exclusions. A technically detailed listing still needs a commercial scope.

• Hardware: exact system name, Superchip count, memory variants, interconnect, host resources, storage, network adapters, racks, and spare parts.
• Software and support: included licenses, subscription duration, support level, renewal charges, and responsibility for updates.
• Deployment: installation, integration, testing, training, and acceptance criteria for the intended workload.
• Facilities: specified power, cooling, rack space, weight, cabling, and any site modifications.
• Commercial terms: taxes, shipping, payment schedule, warranty, hardware condition, and cancellation terms.
• Availability: the actual delivery commitment and what happens if the proposed configuration changes.

Ask the supplier to separate one-time costs from recurring charges. If support or software is bundled, establish what happens after the included period. Do not apply generic percentages for maintenance, depreciation, or installation to a system whose configuration and contract are still unknown.

Build a cost comparison around completed work

For ownership, include acquisition, deployment, facility changes, electricity, cooling, storage, networking, licenses, maintenance, and operating staff over the evaluation period. Record utilization assumptions and expected downtime. Treat any resale value as an explicit assumption.

For rental, include the full instance allocation, runtime, storage, transfers, software charges, setup work, idle resources, and failed runs. Confirm whether the price is per GPU, Superchip, server, or cluster. A discounted rate may depend on a commitment or an interruption policy.

Cost per completed job = total attributable cost ÷ successfully completed jobs. Use equivalent output-quality and completion criteria on both sides. Comparing a supercomputer purchase with centuries of single-GPU rental does not establish savings when the two configurations complete different work at different speeds.

A hypothetical scaling example

Suppose a job takes 12 hours on four GPUs billed at €2 per GPU-hour: compute costs €96. If eight GPUs finish the same validated job in eight hours at the same rate, compute costs €128. The larger configuration finishes four hours earlier but costs about 33% more for that run.

These are illustrative inputs, not DGX GH200 or provider prices. They show why GPU-hours, elapsed time, and cost must be measured together. Add setup, storage, licenses, and failures before making a purchase decision.

When a smaller configuration is worth testing

Start with the smallest configuration that can meet the application's memory, software, and performance requirements. Independent experiments, supported rendering jobs, or inference that fits on one GPU may not need a large shared-memory system. A workload with frequent cross-GPU communication needs a different evaluation.

For that smaller-workload path, the Hivenet rental page advertised RTX 5090 instances with 32 GB of VRAM from €0.75 per hour when checked on October 16, 2026. Confirm the active preset, capacity, and rate before budgeting. Compute with Hivenet uses prepaid credits and per-second billing while instances run.

That starting rate is not a DGX GH200-equivalent price. An RTX instance does not reproduce the system's memory capacity or NVLink Switch topology. Benchmark a supported workload and include the resources needed for a usable result; a lower hourly rate alone is not evidence of lower total cost.

Questions to settle before purchasing or renting

Is there a verified public NVIDIA DGX GH200 list price?

We found no public list price in the NVIDIA product material checked for this guide. Obtain a current quote that explicitly names DGX GH200 and lists the complete configuration. A price without that scope is insufficient for a purchase budget.

Is GH200 the same as H200?

No. GH200 combines a Grace CPU and Hopper GPU. H200 is a GPU product. Check the exact server architecture, memory, and accelerator count rather than relying on the similar names.

Does renting GH200 hardware provide a full DGX GH200 system?

Not necessarily. Ask what the allocation contains and how its processors communicate. A provider may supply one Superchip or a smaller server without the complete DGX GH200 fabric.

What is the next practical step?

Write a representative workload specification, obtain matching hardware and rental proposals, and test the configurations that can actually run it. Choose from measured time, cost, output validity, and operational requirements. Leave the purchase price open until a supplier provides a quote you can inspect.